Predicted LTV (pLTV): what it is and when to bid on it

Predicted LTV, often shortened to pLTV, is an estimate of what a customer will be worth in a set window, such as the next 12 months. “pLTV” on its own also turns up in baseball, so in this guide it always means the ecommerce metric.

It is written for PPC specialists and ecommerce owners who bid for new customers on Google Shopping and want to know how much faith to put in a number that hasn’t happened yet.

How is predicted LTV different from observed LTV?

Observed LTV is the revenue or profit a customer has produced so far, and predicted LTV is an estimate of what they will produce in a window that has not ended. The first is counted from orders, and the second has to be checked against them later.

Observed LTV includes order revisions: a refunded or cancelled order leaves the figure. It is the number behind LTV:CAC for ecommerce, and it can be counted from orders without any forecast.

Both numbers need a window and a measure. “LTV” alone does not say whether it is revenue or profit, or whether it covers 3, 12 or 24 months, so a prediction is only comparable with an observed figure measured the same way. Profit is the better measure for bidding, because two customers with the same revenue can leave very different profit.

How do you estimate predicted LTV?

You can estimate LTV from a historical average per customer, from a cohort curve built on past orders, or with a statistical or machine-learning model. The first two are arithmetic you can do in a spreadsheet, and each one breaks in a different place.

The hardest customer to estimate is the one with a single order, which is every new customer at bid time. Nothing about their behaviour after the first order exists yet, so any method gives them the average of similar customers. Only a model that uses what is known at the first order, such as the product bought, can say more, and it needs plenty of past orders to learn from.

A cohort curve is the easiest to explain. If a cohort placed 1.30 orders per customer by month 3, you can extend that pace and say what profit per customer it implies at month 12. Your repeat purchase rate is the main input to it.

Small shops are where all three struggle. A handful of large buyers can move an average a long way, which is why an estimate built on 20 customers deserves less trust than one built on 2,000.

One cohort, counted then extended

In practice the observed part of a customer’s value comes first and the predicted part is added to it. The chart below follows one cohort: the first months are counted, and the later months are an extension of that pace.

For example, a customer’s first order brings €40.00 of profit. By month 3 the cohort has placed 1.30 orders per customer, so profit is €52.00. If the pace holds, orders reach 2.20 per customer by month 12, which is €88.00.

The dashed line is a target, here an acquisition cost (nCAC) of €60.00. The prediction says the cohort repays it between month 3 and month 6. That statement is a forecast until month 6 arrives. The over-estimate section below shows what happens when it turns out wrong, and customer acquisition cost for ecommerce explains how to work out nCAC. The MER and nCAC calculator gives your own figure from ad spend and new customers.

Why bid on a prediction at all

When Smart Bidding places a bid, the customer has not bought again, so the first order is all there is to value. A prediction lets you value that buyer for what they may do next instead of for the basket alone.

Without a prediction you either bid on the first order’s profit alone, or you wait months for observed value. Google Ads Help describes the trade-off: “Short-term conversion values can be useful when you want to maximize immediate profit or customer acquisition as cash flows allow. Lifetime conversion values can be more useful when trying to maximize long term growth.”

Google also says that, when you can’t track a value definitively, “using a 15-20% conservative estimate is often more helpful than using no value at all.” That sentence is about unmeasured gains such as word of mouth, and not about LTV. The principle carries over: a cautious estimate you can defend beats no value, so shade a prediction down before you bid on it.

What an over-estimate costs

If a prediction is too high, you pay too much for every new customer it covers, and you only find out when the window ends. The cost grows with the gap between predicted and observed profit.

Take the cohort above. At month 3 the prediction says €88.00 by month 12. The cohort then orders less often, and reaches €70.00. The gap is €18.00 per customer.

Suppose you want to keep half of each customer’s profit after paying to win them. Then the most you can pay is half the value you used.

The first order alone supports €20.00, the prediction €44.00 and the observed result €35.00. Paying on the prediction gave away €9.00 per customer. Checking it takes a few steps, and LTV:CAC for ecommerce shows how to compare the result with nCAC.

Do Google’s lifecycle goals predict LTV?

No. Google’s customer lifecycle goals tell Smart Bidding to treat new customers differently. Google Ads Help says they “help you increase value from both new and existing customers”, and lists them for Performance Max, Search, Shopping and Demand Gen campaigns.

The acquisition goal New Customer Value is described as “Bid higher for new customers than existing ones”. That is the adjustment the metric in the sidenote above reports: a higher value on first purchases from new customers. Whatever its name suggests, it looks no further ahead than the first basket.

Prediction is optional; observed profit is the safer place to start, and it is one of the conversion actions Product Metrics sends to Google Ads, alongside new customers, returning customers and lifetime value. For the break-even floor any value starts from, use the break-even ROAS calculator, and see POAS vs ROAS for the profit version.

Observed first, prediction second

Start with observed profit per customer by first-order month, so you know what your cohorts produce. Only then decide whether a prediction would add something, and write its window and measure down before you use it.

When the window ends, compare prediction and observed profit, and lower the next prediction if it was too high.